The Transformation of Research and Development Processes through Artificial Intelligence
Definition and Scope of Artificial Intelligence
Artificial intelligence refers to a branch of computer science in which machines are taught to imitate human cognitive functions. As such, there are various methods and techniques to create AI-based systems, including machine learning, symbolic reasoning, search methods, neural networks, and others.
The development of such systems is grounded in several rather diverse scientific disciplines, such as the philosophy of mind, neuroscience, cognitive psychology, computer science, parts of the field of linguistics, and many others.
AI may focus on, among others, learning from data, processing natural language phenomena, creating machines that understand the content of images and videos, simulation of the human brain and natural intelligence processes, and others.
By using similar principles, it is possible to design and develop intelligent systems that are capable of supporting and augmenting performance in various human activity fields, such as production, planning, scheduling, diagnosis, identification, forecasting, prediction, and others.
This multi-paradigm approach toward AI systems allows the relaying of content knowledge so that it can be transformed and used for universal support in research and development issues, regardless of their scientific background and subject domain.
Within this system, the possibility exists for one specific AI paradigm to provide support for research on various subjects simultaneously and for the application of similar content-required solutions across several specific research sectors.
As a result, AI is considered transformative in this domain; do not treat AI as an ordinary tool, as its deployment implies a definite objective and technological transformation in the research process that is above the head. To this day, it has been difficult to create a precise, unique definition of AI due to its complexity, the evolution of its applications, and the lack of agreed-upon definitions and principles of its development; its extended range of applications and usage further complicates the creation of a unique definition.
Historical Development and Evolution
Drawing on its origins in research devoted to a general theory of learning and the simulation by machines of the intellectual faculties associated with reasoning, perception, and planning that had been considered a prerogative of humans and animals, AI has matured significantly, again in parallel with progress in computing, which is fundamental to tackle questions combining the understanding of complex systems and the designing of processes, products, and services.
Below, we try to sketch a historical and thematic perspective that goes from the speculative searches of the mid-twentieth century to the pragmatic implementation of the additional support to human complex activities that is now commonly intended as AI, or rather as AI already located close to, and often overlapping with their natural environment, in the many examples of machine learning and data-driven systems applications we are all more or less familiar with.
This journey covers several key issues, such as: the algorithms guiding the basic capacity of any computer to draw automatic responses or identify and optimize potential strategies, the resources made available to them and now dramatically conditioned by the enormously growing input of data and the increasing power of computing, the models underlying recognition whether of familiar objects or of the novelty of different situations, and of planning.
In the late 1950s, the idea of "information technology" was introduced, which asserted that computers would significantly disrupt the economy by destroying jobs and simultaneously creating demand for those who could exploit new technological opportunities.
As the variety, availability, and, more recently, the applicational complexity of computational resources progressively increased, machines have instead largely supported scientific work, also as a champion for easing routine, repetitive, and redundant tasks, for example, bibliometric analyses, whose application has made available simple and ample resources for identifying interesting research questions, fields, connections, and novel topics.
Moreover, in addition to producing standard bibliometric data, software could thus trace, visualize, and analyze a relative geography of technology, identifying research buildings and districts in their excellence and correlation to the production of specific scientific and technological results.
Such technology was therefore able to provide standard data relevant to describing technical and thematic research evolution, individual behaviors, and interdisciplinary connections; to make available tools to support multipart eye-scan time clustering technologies more broadly described in part one as patent scanning, and in this section as part of general thematic optimization sophisticated institutions profiles plagiarisms with usage compliance sheets.
Based on the current state of the art of AI, our society's need for moving knowledge forward, acting on problems and questions on which existing research communities have built rather solid foundations, has decreased, as AI technologies allow scientists to discover potential connections that are not intuitive from a solely human perspective.
For example, it is now common to hear and use tools that represent recommendation systems to inform the referees' selection for reviewing a scientific paper by checking two criteria: avoiding the selection of a person who has declared a conflict of interest or representing a senior or a junior tutor who can fairly evaluate the overall metrics of a junior-previous-taught candidate. This applied field of AI supports management and first decision-making for the researchers at large, based on the presumption of a fair use of such data-checked.
Applications of Artificial Intelligence in Research and Development
Artificial intelligence (AI) can streamline, revolutionize, and sometimes largely replace various R&D (research and development) activities across life and material sciences. The field of drug discovery is one of the primary research areas where AI has been systematically disrupting traditional R&D activities.
Using AI technologies in drug discovery can save billions of dollars in terms of the length of projects and costs, erasing many drug candidates before reaching clinical phases—expediting the overall development time and reducing the number of compounds to test.
Projects based on AI models in drug development due to linked pharmaco-medical big data can last around 3–12 months. In some cases, leading pharmaceutical companies have used AI projects in combination with their molecular properties of virtually designed compounds.
Compounds become 1–9 times more active or induce an estimated 10% improvement in treatment efficacy. AI has influenced traditional R&D activities further than drug discovery.
Today, many AI-based models and algorithms support materials science, forensic sciences, and research projects in physics. AI is widely used in predictive data analysis and numerical simulations of materials and medical devices in life science projects.
This can identify new research strategies and data characteristics. Predictive AI technologies also support nanodevice research and 3D printing by identifying eligible collaborators and project-relevant literature.
Thus, data-intensive research projects in materials science or medical devices can also benefit from specialized AI-based collaborative tools based on learning from public databases.
In conclusion, in combination with learning, AI can integrate intelligence experiments in the CRISPR research portfolio, predictive data analysis, image analysis, and other big data modalities as defined earlier. However, the field remains in development, and governing policies for learning and adaptation are still evolving. Even so, many new discoveries in interdisciplinary research can be expected.
Drug Discovery and Development
Artificial Intelligence (AI) has gained increasing attention for its utility in accelerating drug discovery and development, particularly in the ability to process large amounts of chemical data to identify possible drug candidates in a quicker and more efficient manner than traditional methods.
Machine learning (ML) algorithms can be used to predict either molecular interactions between active substances and proteins, metabolic and toxic effects in the body after administration, or biological responses to a medical treatment.
Together with data warehousing, ML can accelerate the drug design process by providing materials for screening while providing a deeper level of drug moiety property prediction as well. In the clinical phase of drug development, AI can be used for clinical trial design optimization, as well as improving patient selection.
This modification of trial design and patient selection can decrease development-related costs as well as increase the success rate of the therapy in terms of beneficial effect towards probability without adverse effects due to findings with smaller sample sizes. In clinical trial design optimization, data from various resources can be applied to develop prediction models for patient responses which will be extrapolated for prospective investigational trials, resulting in more efficient tools for appraising the drug or drug combination.
Successful AI applications in pharmaceutical research and development include the repurposing of existing medications for other anti-inflammatory conditions, and the exploitation of existing drugs in new therapeutic contexts, both of whom discovered that existing drugs could impede novel host-pathogen interactions.
A recent review provides an inclusive overview of AI-based drug discovery. Although this disruptive technology may accelerate drug development, several ethical bottlenecks have to be considered before beginning with AI-based patient therapy, such as the need to guarantee the safety of the patient and to find innovative and adaptive approaches to regulate these new agents.
In addition to bringing efficacious drugs to the market faster and more accurately, multimedia analytics remain a very active discipline with the possibility of unveiling the underlying nuances of multi-domain correlates and identifying novel patterns to optimize drug delivery, improve patient care, spawn new types of treatments, broadly identify room for improvement in healthcare systems, build predictive models, and more.
The fast pace of change in deep learning and multimodal methods further simplifies integrating different kinds of data in an adaptive learning approach. Although changing rapidly, synthetic intelligence (AI) has, is, and will revolutionize many aspects of human existence, including the healthcare and pharmaceutical landscape.
Recent advances in AI abilities to analyze and interpret large data points are becoming pivotal in uncovering early signs and contributing to the discovery, treatment, and understanding of aberrant biological and cognitive states.
With future AI or AI-inspired healthcare initiatives, extensive possibilities can be expected in personalized medicine, preventive healthcare, drug discovery, development, and repositioning.
The analysis of multimodal data, including but not limited to MRI/fMRI, PET/SPECT, laboratories, clinical assessments, electrophysiological recordings, activity tracking, and other environmental data, has shed new light on mental illness.
Although AI/ML might be used most immediately for large dataset integrations in these human studies, particularly for building early intervention predictive models to guide when and to whom targeted therapies may be helpful, foundational multimodal AI/ML advancements in connectomics and bioinformatics research have been extensive and are certainly within the scope of this special issue.
Materials Science and Engineering
Artificial intelligence (AI) has the potential to revolutionize materials science and engineering. Utilizing predictive modeling for experimental synthesis and properties and using these to optimize predicted performance properties with extensive parametric analysis, one can identify material systems useful for unique applications.
The development of AI, and in particular machine learning (ML), has opened a new era in the utilization of these tools for this type of modeling. An integrated experimental and machine learning approach to addressing both of these prediction challenges enables collaborative access to large databases of material properties constructed by iteratively populating a Materials Genome Database by predicting new sets of materials to synthesize and assess.
These tools allow analysis across trends in experimentally derived materials properties to gain insights into the behavior of materials and mechanisms that control their performance.
There are numerous examples illustrating the value of AI in materials science and a clear path forward toward broader materials impact. For example, a combinatorial approach using AI has allowed the discovery of new materials in the electronics sector that move AI from simply allowing data prediction to contributing to the optimization process for material design.
Other high-impact studies address nanoscale effects, which are costly to study in experiments, and accelerating the use of biomaterials with AI. Looking forward, it may also be possible to employ AI to mitigate the environmental impacts of the large-scale production of new materials by multiple industries for new engineering solutions. AI developments will continue to extend the scope of materials that can be modeled.
Challenges and Ethical Considerations in AI-Driven R&D
The widespread application of AI drives not only the digital transformation in the R&D sector; it also radically changes the processes of developing new products, services, and ultimately, the ways in which scientific results and technical problem solutions are established and validated.
AI-based processes feature new stakeholders and capabilities that need to be integrated, while established divisions of labor are called into question. Important accompanying topics include changes in the workplace, ethical and legal challenges in science, the transformation of research infrastructures, and fair access to opportunities related to AI.
For many years, the research data center has been involved in the development of measures that allow the potential of AI in data linkage to be realized. An essential part of this work is to consider the quality, security, infrastructure, documentation, and storage of data.
In introducing AI into data-driven R&D processes, and in designing a framework for integrating different evolving AI-based technologies for specific, often highly sensitive, research areas, there are five slightly overlapping challenges to be addressed:
- Regulation and Ethics,
- Bias in AI-assisted results,
- The realization of AI-based insights: inscrutability and trust,
- Strategic and technical collaboration, and
- Standardization and sustainability.
Data Privacy and Security
In AI-driven R&D, its benefits and challenges go hand in hand. While AI can be used to find needles in the haystack, from a data privacy and security point of view, haystacks represent valuable collections of patient and healthcare data generated from years and sometimes decades of research and development at immense costs, not only of money but also of time and ethical concerns.
The development of regulations and ethical frameworks around the world safeguards the collection and processing of such data. Fundamental to this approach is the concept of data stewardship, ensuring appropriate physical data security and privacy protection, which is generally required by law and/or enforced as a guideline by funders and institutions.
Data can be processed and shared with a view to privacy, on grounds of scientific investigatory purposes, but only where strict technical and organizational constraints have been met. The sharing of data must respect privacy, including informed consent of participants, the optional withdrawal of a participant at any point (without reason or any effects), and a fundamental assurance that data will have been collected on a need-to-know basis.
In other words, data will be processed to the minimum extent possible to achieve the primary purpose of the data.
Balancing the need to train AI with data from the reality it purports to represent, the need for privacy comes into conflict with the need for AI to learn from the most data possible. Technologically, different solutions could address these privacy issues directly.
The concept of “dilution” works on the basis of anonymization; simply, reduce detail such that data cannot be re-identified. Such an approach is widely defined, but where it functions effectively for individual private citizens, the practical application in AI is cause to limit the data’s utility for learning. The critical driver for AI is, of course, the retention of classification capacity to be effective.
Encryption too would be an effective solution, with differential privacy or secure multiparty computation promising a future basis for this approach, but considering recent trends in AI technology, encryption seems to overly limit the utility of the data value needing scientific research breadth in AI augmentation of scientific data.
Misuse of patient data to train AI that leads to misclassification or other problematic outcomes can lead to significant damage in patient and scientific enterprises, amplified by media and political outcries.
Limiting sharing or use of such data unnecessarily, in short, hampers scientific progress and can incite justified distrust in the application of AI in general. There is increasing discussion in the AI space on the development of ethical AI, rather than “good AI,” which is more about instilling trust between man and machine in a scientific process.
Key to this is the fallback of research ethics and data practices that are common everyday points of reference: data transparency, the availability and application of consent, the avoidance of research involving vulnerable groups, and avoiding the damage or disrespect of research participants. All these ideas are heavily interlinked and revolve around open and transparent responsible AI uptake.
Bias and Fairness in AI Algorithms
Bias and fairness within AI algorithms have been critical issues in the recent past. Biases can be introduced at several stages of AI models used in research and development processes, from data collection to training and evaluation, which may inadvertently result in biased research outcomes.
This is particularly vital in research and development initiatives that use machine learning and AI-driven insights, as biased results can disproportionately direct decisions affecting the studied phenomenon and potentially have additional societal impacts.
When leveraging AI models for research and development, it is vital to discuss and anticipate possible biases in research outcomes that may result from the use of such algorithms. A consideration of this issue and transparency in model building and algorithmic assumptions can incite dialogues that root out these biases and ensure fair information conclusions.
Because of the importance of this issue, multiple efforts are underway to address biases in AI algorithms. The inclusive database includes multimodal signals obtained from real-life conversations between native English and non-native English speakers and designed to be used for automatic detection of non-native speech.
Not all communities, access to information sources, or organizations are equally equipped to develop models with fair outcomes. This substantiates the need for varied datasets of inclusively collected data across communities or datasets developed and collected by responsible organizations.
There are several efforts to automate the assessment of algorithmic fairness from the perspective of datasets and training. All these initiatives suggest a diversity of researchers and stakeholders engaging in the creation of an AI algorithm to address the potential social concerns and rebalance the differences highlighted by current datasets. At the same time, these accentuate the importance of public inclusivity and democratic values in the development of such algorithms.
Future research directions include the development of specific metrics to ensure the fairness of AI models generating research and development insights, evaluative frameworks, and accentuating the necessity for research and development processes to remain vigilant of potential biases and choose fairness above all.
Future Trends and Opportunities in AI for R&D
Emerging trends and developments within AI are a focus of increasing attention, including algorithmic strides such as self-supervised learning and more powerful deep learning frameworks, as well as hardware shifts such as the increasing availability of dedicated neural processing units.
In the context of R&D, where time to market is often longer and scientific principles limit rates of advancement, opportunities for AI in basic research are only beginning to emerge, with the increased experience and skill of AI experts around the world anticipated to drive a proliferation of more interdisciplinary projects.
Opportunities include better predictive performance of molecular properties by machine learning on quantum chemistry data, as well as new approaches in areas where the limitations of traditional reductionist methods are apparent, such as in the use of machine learning derived interpretable models to personalize drug combination treatments.
Future trends include advances in Explainable AI (XAI) and systems that interweave the development of new hypotheses with automated experiments in a way that learns from both, as well as the adoption of adaptive learning systems that improve based on use and feedback. Research is needed to further explore and articulate these technical possibilities as part of a broader societal conversation.
As demand in fields of applications such as healthcare rises, there is potential for more controversy, conflict, and impact as AI systems move from hypothesis generation to deployment and decision making. XAI in particular is growing in importance as a subfield of research that seeks to develop methods for improving the explanations of decisions and reasoning processes that underpin AI systems, to help foster intent trust in or adherence to the outputs or predictions of AI systems.
The investment in ethical design, human AI interactivity, and XAI is also expected to increase as proposed legislation will, if enacted, require all high-risk AI systems to be designed to produce detailed human interpretable logs.
The ability to practice “good science” while using AI methods that are inherently more complex can therefore be seen as a future competitive advantage with added value. This is the argument behind creating adaptable systems that learn from interactions with knowledgeable humans or users over time, instead of being locked in at the time of deployment.
Promising potentiators of AI systems in R&D include human in the loop, multi-armed bandit, and reinforcement learning control strategies for experiment selection under uncertainty. In seeking to produce robust adaptive systems, a key potential challenge will be a cultural divide between the experimental method optimization and digital domain communities, as well as potential regulatory, safeguarding, and cybersecurity concerns regarding human and machine interaction. Such systems can already be legally used, but widespread application would benefit from greater ethical engagement upfront.
Explainable AI and Interpretability
Explainable AI (XAI) is a rapidly rising theme of research and development aimed at making AI systems interpretable to their users. Key stakeholders and decision-makers use explanations to understand an AI model and are knowledgeable about the predictions to make certain decisions. Interpretability is the act of explaining a model’s decisions.
There are various methodological approaches to explaining a model, some of which divide AI frameworks accordingly, such as model interpretability techniques and separately developed AI explanation methods, as well as visualization tools for different types of AI frameworks.
The levels of methods can range from global to local explanations using either feature attributions or concept-based methods.
Explainability links directly with the concept of responsible data, in that should future biases be discovered or other ethical implications arise, researchers and developers are able to explain this to their stakeholders and address it accordingly. It draws together XAI, Interpretable Machine Learning (IML), bias and fairness, and the requirements of policymakers, sectors, and society to provide AI explanations. In medical risk assessment and software system design, it is used where policies and compliance standards require interpretability.
Therefore, to enhance compliance and trust, methods of AI explainability are critical. Methodologies for providing XAI tools and visualizations are a rapidly growing area of research that needs continuous attention and development.
As AI is changing rapidly, new methods and AI frameworks mean that it is difficult for XAI methods to keep up. In a survey carried out by many companies, industries, and countries in the European Union selected for AI research and standardization, many of the selected questions had very high consensus on the critical role of XAI, and several were about what needed to be addressed to achieve this. It is therefore assumed that the European Union and its companies consider XAI important for ensuring AI accountability and developing advanced research into healthcare AI.